Managed 300+ tickets per quarter and enforced a formal escalation playbook, reducing median first-response time below 1 hour and recording 0% logo churn across all EMEA enterprise accounts.

Managed >200 Mobileye engineers via onboarding, workflow building, and proof-of-value workstreams, delivering a $1.5M three-year growth agreement.

Authored and maintained a reference repo for a full-fledged SageMaker MLOps pipeline, from experiment tracking to production monitoring, now used by Fortune 500 R&D teams.

Provided a competitive overview (Weights & Biases, MLflow, Databricks) and conducted enablement workshops for 20+ marketing & sales colleagues, optimizing positioning and closing cycles.

Implemented Azure OpenAI, Azure AI Search (Cognitive Search), AKS/Functions, Event Hub, and Cosmos/Redis integrations.

Delivered secure VNet patterns, key‑rotation, cost dashboards, and blue/green rollouts for LLM services.

Partnered with enterprise clients to design end‑to‑end evaluation pipelines using Comet Opik.

Defined taxonomies (hallucination, grounding, refusal/stance, PII/toxicity, compliance), sampling strategies, human‑in‑the‑loop review, and CI quality gates wired into Azure DevOps/GitHub Actions.

Implemented a Deep Learning model serving platform using Torch Serve on top of Azure Kubernetes Service, cutting development time by 2 months.

Fine-tuned Torch Serve parameters through load testing for the image embedder model, increasing throughput by 40% and reducing response time by 60%.

Designed and conceived a scalable microservices architecture with automated CI/CD flow, reducing development and deployment efforts by 80% for future services.

Diagnosed slow response times for a similar items service using profiling, reducing response times by 60% after optimization.

Operated Elasticsearch tuning efforts, reducing initial response times for the worst case from 11 to 3.5 seconds.

Created and took full ownership of a scalable scraping architecture, allowing to keep the pace with eCommerce marketplace page changes without changing the codebase, reducing development time and maintenance by 70%.

Improved scraping velocity and number of page scraped per hour by 60%.

Developed UI for the dynamic scraper using React from the ground up.

Implemented ~2,000 lines of code, pushed 80+ commits, and resolved 30+ bugs.

Built a "model-as-a-service" solution on Azure AKS and implemented K8s + Flask stack and GitLab CI pipelines that enable data-science notebooks to be production endpoints within < 1 hour (down from days).

Replaced a legacy C++ image-tagging model with a TensorFlow/Keras implementation served on GPUs (precision ↑ 11% and codebase ↓ 80% LOC) and integrated the new API directly into an Erlang back end after learning the language.

Created a scalable training pipeline by deploying Apache Airflow on AWS EKS with shared DAGs stored on EFS, eliminating laptop-based executions and decreasing experiment turnaround time by ~60%.

Documented end-to-end SOPs for researchers and back-end engineers, standardizing hand-offs from research, deploying and minimizing integration friction.

Upgraded device-insight API and re-architected an AsyncIO + SQLAlchemy Python service offering device metadata to analytics teams, cutting average query latency by ~45% and doubling daily query rate.

Built a high-volume device-classification pipeline and worked with data scientists to ramp up a Pandas prototype into a PySpark workflow on AWS EMR, processing 10 M+ devices per run and reducing ETL costs by ~30%.

Supported real-time forecastability and ran a Kafka-streaming job on Kubernetes that scores tens of thousands of new devices per second, returning results into Armis's core platform in sub-second latency.

Scaled up for performance and optimized Spark joins, partitioning, and cluster autoscaling, decreasing end-to-end runtime from hours to less than 25 minutes.

Worked across departments, acting as an intermediary between data science, back-end, and DevOps teams, ensuring smooth hand-offs from research to production and developing extensive run-books for future maintenance.